Darwin-WGA. A Co-processor Provides Increased Sensitivity in Whole Genome Alignments with High Speedup

Transcription

1 Darwin-WGA A Co-processor Provides Increased Sensitivity in Whole Genome Alignments with High Speedup Yatish Turakhia*, Sneha D. Goenka*, Prof. Gill Bejerano, Prof. William J. Dally * Equal contribution

2 What are whole genome alignments (WGA)? 2

3 WGA is correspondence between genomes For each segment of the target genome, corresponding segments are found in the query genome Match Deletion Insertion Mismatch Indel = Insertion + Deletion Query Chimpanzee genome Chr5 Target Human genome Chr5 Cabanettes et al.,

4 Why are whole genome alignments important? 4

5 WGA help predict functional elements Control logic for gene expression Enhancer gene Protein recipes that encode function Regions with sequence conservation (Mayor et al., 2000) 5

6 Classical alignment algorithm 6

7 Smith-Waterman Algorithm Inputs Target sequence (r) GTGTCACTA (L r = 9) Query sequence (q) GGCCAACTA (L q = 9) Scoring parameters Smith-waterman equations V i, j = max V i 1, j 1 + W(r i,q j ) V i 1, j + gap V i, j 1 + gap 0 Alignment GTGTC-A-CTA G-G-CCAACTA W = Gap penalty = 1 7

8 Smith Waterman algorithm intractable on whole genomes Smith Waterman algorithm time and space complexity ~ O(L r. L q ) Mammalian genomes ~ base-pairs Use heuristics based approaches seed-filter-extend 8

9 Seed-Filter-Extend algorithm (LASTZ) 9

10 Seeding finds small matching local patterns Query GTAGCGGGCGTCTGAC Target GTAGCAGTC GTAGCAGTA GAGCTGCAA Seed hits Seeding finds local matching patterns of fixed length s Substrings of query of length s are compared to the target Substrings start from position 0 Seed 10

11 Seeding finds small matching local patterns Query GTAGCGGGCGTCTGAC Target GTAGCAGTC GTAGCAGTA GAGCTGCAA Seed hits Seeding finds local matching patterns of fixed length s Substrings of query of length s are compared to the target Substrings start from position 0 Seed 11

12 Filtering aligns ~100bp around seed hits Target Calculate scores along the seed hit diagonal (match or mismatch) Query Track maximum score Seed hit Filter (Ungapped) Stop as soon as score falls below (max_score - x) Does not consider indels 12

13 Filtering aligns ~100bp around seed hits Target Calculate scores along the seed hit diagonal (match or mismatch) Query Track maximum score Seed hit Filter (Ungapped) Stop as soon as score falls below (max_score - x) Does not consider indels 13

14 Filtering aligns ~100bp around seed hits Target Calculate scores along the seed hit diagonal (match or mismatch) Query Track maximum score Seed hit Anchor Filter (Ungapped) Stop as soon as score falls below (max_score - x) Does not consider indels 14

15 Extension uses Y-drop algorithm Anchor Query Target Start computing score along a row when score > (max_score - y) Stop computing score along a row when score < (max_score - y) score < (max_score y) Extend 15

16 Extension provides the final alignments Target Query Final alignment = right extension + left extension Anchor Extend 16

17 Why is LASTZ less sensitive? 17

18 Increasing indel frequency => increasing need for gapped filtering -88 Mil yrs -6.4 Mil yrs indel per 625 bp 1 indel per 30 bp LASTZ s filter threshold Replacing ungapped filtering by gapped filtering slows down the software by 200x 18

19 Darwin-WGA algorithm overview 19

20 Seeding D-SOFT Target bin and Query chunk determine diagonal band Each seed hit falls in a single diagonal band At most 1 seed hit per diagonal band is extended Query Reference Target Chunk 1 Chunk 2 Chunk 3 20

21 Gapped Filtering Banded Smith Waterman Seed hit extended using banded Smith-Waterman Pre-determined band with no traceback If (max_score > threshold), the maximum score position (x max ) is the anchor or the starting point for alignment extension Gaps considered => better alignments for species further apart Query Target Seed hit! max 21

22 Extension - GACT-X algorithm Tiled (tile size T, overlap O) implementation inspired by GACT in Darwin* Origin of the next tile lies at the intersection of the current traceback path with the overlap Anchor Staring point for next tile T1 Drop these pointers Outside tile overlap Inside tile overlap On tile overlap border * Turakia et al., ASPLOS 18 22

23 ... Extension - GACT-X algorithm Extension along a direction continues until a tile is encountered with a non-positive maximum score Anchor T1 Right extension T2 Outside tile overlap Inside tile overlap On tile overlap border 23

24 ... Extension - GACT-X algorithm Final alignment combines left and right extension Staring point for next tile... T1 Anchor Left extension Right extension T1 T2 Outside tile overlap Inside tile overlap On tile overlap border 24

25 Extension - GACT-X algorithm Y-drop implementation within each tile Target Adaptive band with traceback Reduces on-chip memory requirement compared to computing whole tile Reduces compute time Anchor T1 Query Outside tile overlap Inside tile overlap On tile overlap border * Turakhia et al., ASPLOS 18 25

26 Workloads in LASTZ v/s Darwin-WGA LASTZ Seeding 13B seed hits Ungapped filtering 300k anchors Extend 150k alignments Darwin-WGA Seeding (D-SOFT) 13B seed hits Gapped filtering 1.2M anchors Extend 700k alignments 26

27 Whole genome alignment v/s Read alignment 27

28 1. WGA requires aligns less similar sequences Genomes may diverge considerably over evolutionary timescales and have low sequence similarity Read alignment deals with highly similar sequences (wellcharacterized sequencing error model) % alignments % similarity in read alignments v/s whole genome alignments % similarity Read alignments Whole genome alignments 28

29 2. WGA has longer alignments with large indels Whole genome alignments can span millions of basepairs with large indels Read alignments span not more than tens of thousands of base-pairs with much shorter indels Previous hardware accelerators would require high on-chip memory percentile alignment lengths in read alignments v/s whole genome alignments 1,000 10, ,000 1,000,000 alignment lengths WGA alignment Read alignment

30 Darwin-WGA Framework Seeding done in software Banded Smith-Waterman and GACT-X accelerated in hardware as bounded Dynamic Programming with Systolic Arrays 30

31 Hardware Acceleration 31

32 Accelerating bounded Dynamic Programming with Systolic Arrays 32

33 Accelerating bounded Dynamic Programming with Systolic Arrays 33

34 Accelerating bounded Dynamic Programming with Systolic Arrays 34

35 Accelerating bounded Dynamic Programming with Systolic Arrays 35

36 Accelerating bounded Dynamic Programming with Systolic Arrays Banded Smith-Waterman - preset band and no traceback GACT-X - adaptive band with traceback 36

37 Evaluation Framework 37

38 Experimental Setup CPU (baseline) AWS c4.8xlarge instance 36 vcpus (18 physical cores) LASTZ as software baseline Parasail to estimate isosensitive runtime $1.59/hour FPGA (Darwin-WGA) AWS f1.2xlarge instance 1 Xilinx Virtex Ultrascale+ FPGA (50 BSW and 2 GACT- X arrays with 32PEs) 8 vcpus $1.65/hour 38

39 ASIC TSMC 40nm DC synthesis (not a chip prototype) Configuration Area (mm 2 ) Power (W) BSW Logic 64 x (64PE array) GACT-X Logic 12 x (64PE array) Traceback SRAM 12 x (64PE x 16KB/PE) DRAM DDR4-2400R 4 x 32GB TOTAL Power (in W) CPU FPGA ASIC 39

40 Species and Genome Assembly Target Species C. elegans (ce11) D. melanogaster (dm6) Size (Mbp) Query Species 100 C. briggsae (cb4) 137 D. simulans (drosim1) D. Yakuba (droyak2) D. pseudoobscura (dp4) Size (Mbp) C. elegans 0.51 C. Briggsae Roundworm family D. melanogaster 0.06 D. simulans 0.05 D. Yakuba 0.11 D. pseudoobscura Fruit fly family dm6-drosim1 molecular distance comparable to human-monkey dm6-dp4 molecular distance comparable to human-chicken 40

41 Results 41

42 Darwin-WGA finds genes that LASTZ does not Seed hit 1 Seed hit 2 Seed hit 3 Indels (shown by arrows) around each seed hit dropped by ungapped filtering (LASTZ) but retained by gapped filtering (Darwin-WGA) 42

43 Darwin-WGA Sensitivity Improvement Versus LASTZ Species pair Top-10 Alignment Chain Scores Matching Base-pairs within Alignments Number of Aligning Exons (protein-coding genes) dm6-drosim % 1.25x +0.20% dm6-droyak % 1.41x +0.09% dm6-dp % 1.42x +0.41% ce11-cb % 3.12x +2.70% Represent orthologous sequences (derived from speciation ) Represent paralagous sequences (derived from duplication ) False positive rate (2-mer shuffled genome): % Represent functionally relevant orthologous sequences, under some selective pressure (at least in the target species) 43

44 Runtime and Cost Comparison Species pair LASTZ runtime (sec) Isosensitive s/w runtime (sec) Darwin-WGA runtime (sec) FPGA ASIC FPGA (Perf/$) Darwin-WGA Improvement ASIC (Perf/W) ce11-cb ,960 3, x 1,478x dm6- drosim1 dm6- droyak ,627 5, x 1,547x ,454 6, x 1,539x dm6-dp ,700 4, x 1,553x 200x slowdown 22x speedup 306x speedup 44

45 GACT-X uses 3x less space and time as compared to GACT 45

46 Summary Darwin-WGA replaces ungapped filtering in LASTZ by Banded Smith- Waterman algorithm for higher sensitivity up to 3x matching base-pairs up to 5.7% more orthologs up to 2.1% more exons Darwin-WGA outperforms iso-sensitive software FPGA: 24x performance/$ improvement ASIC: 1,500x performance/watt improvement GACT-X provides 3x improvement in speed and storage efficiency compared to GACT 46

47 Thank You! 47